Re: 5950X or 14900K?

On 1/5/2024 2:30 AM, David Chmelik wrote:
> I have a 5950X workstation, but would a 14900K one be an upgrade or 
> downgrade for computer programming/science code building/compiling/making 
> speed?  Won't use newer AMD until there's a full or larger ATX system-/
> logic-/mother-/main-board with plain PCI (currently only on micro-ATX).
> 

Requires a miracle to get a PCI slot now.
Even though it is a one-chip solution to have a PCI slot (PCIe to PCI bridge chip).

*******

First we check Ark on ark.intel.com .

https://ark.intel.com/content/www/us/en/ark/products/236773/intel-core-i9-processor-14900k-36m-cache-up-to-6-00-ghz.html

    P&E cores, Turbo, 253 watts      # Hefty load

Many programs do not benefit from multiple cores. The
basis of snappy desktop performance, is the behavior of
a single thread task.

https://www.cpubenchmark.net/singleThread.html

    14900K    4,798
     5950X    3,466

Multithreaded.

https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i9-14900K&id=5717

    14900K    61379

https://www.cpubenchmark.net/cpu_lookup.php?cpu=AMD+Ryzen+9+5950X&id=3862

     5950X    45780

The Intel is asymmetric because of P&E.  16+16+36MB=68MB cache
                                         L3 one blob, multiple ports ?

Cache L1: 	80 KB (per core)
Cache L2: 	2 MB (per core)          P-Cores: 8  (x2 core count for Hyperthreaded)
Cache L3: 	36 MB (shared)

E-Core L1: 	96 KB (per core)         E-Cores: 16 (four cores per module)
E-Core L2: 	4 MB (per module)

https://www.techpowerup.com/review/intel-core-i9-14900k/2.html

   Die shot

   https://www.techpowerup.com/review/intel-core-i9-14900k/images/arch1.jpg

   Die shot labels

   https://www.techpowerup.com/review/intel-core-i9-14900k/images/arch2.jpg

"Ring bus" is not the best. On my 4930K, you get five cores performance
from a six core processor, because of "Ring bus". Unfortunately, I have
been unable to find a scaling test that increases core count one at a time.

   Cinebench result here???  Can't find a methodical scaling test web page.

   # Ring bus apparently has options to downclock. Tiny effect on program launch ?

   https://forums.guru3d.com/threads/intel-ring-bus-oc-improvement-to-pc-responsiveness.447118/

Intel had done "mesh" as a technical improvement over it's previous
series-connected ring buses. No mention of mesh bus on this one.
The ring structure is unstated (some large 28 core CPUs used to have three rings
joined at edges).

Now the L3 is labeled as "shared", but is it "segmented" ?

On the 5950X, there are two silicon dies, and each one has L3, and
to maintain their state, in cross-town traffic, would require snoop
traffic. This means on the 5950X, it may have more L3, but the L3 is
"segmented".

On the 5950X, you can take it out of dual channel mode, and
run the two channels separately. That's shown in your
"AMD Ryzen Master" application output. I've not investigated that,
like whether I can even find the setting in the BIOS screen.

   [Picture]

    https://i.postimg.cc/prKSTdqr/AMD-Ryzen-Master.gif

DDR5 memory, a DIMM has two channels on it. A dual channel Intel board
with DDR5 then, counts as four channels. But whether it can be
uncoupled, I don't know if that is possible. Historically, the
industry has been "a bit bus-dishonest", in that a quad channel
board in the past, really only fetched on two channels at
a time, and behaved like a dual channel board. Although there
have been diagrams showing mesh bus plus two reservation stations
for a quad channel board, it's not clear that the design chooses to
use two reservation stations (dual channel per each) at exactly the
same time.

Here, someone has cranked the living shit out of the memory. Notice
how the L3 is only a little faster than main memory.

https://rog-forum.asus.com/t5/image/serverpage/image-id/85105iFE78CC481DEF6913/image-size/large/is-moderation-mode/true?v=v2&px=999

   5764 MHz                                         14900K
   L1       80K  645.5 GB/sec
   L2        2M  137.7         <=== 2M per E-core
   L3       36M   55.1
   DDR5-8200      34.6

A 7950X, same thing. L3 not a big win necessarily. But good speed, whatever the hell they've done to it.

https://www.techpowerup.com/forums/attachments/img_2615-jpg.288099/

   4500 MHz
   L1       64K  302.1 GB/sec
   L2        1M  127.4
   L3       64M   72.5
   Not sure       55.4

My lowly 5950x. Stock clock. Memtest does not seem to turbo?

    [Picture]

     https://i.postimg.cc/MZ7TgJSK/memtest-5950x.gif

   3400 MHz
   L1       32K  292 GB/sec
   L2      512K  139
   L3       64M   48.3
   DDR4-3200      15.8

Since memtest runs on multiple cores now, there are occasional
memory bandwidth reporting errors (factor of 2). The AMD 5600G
can report "40GB/sec" for memory, if you run it XMP profile 2 (=CR1)
and the memory test process uses multiple cores. Which suggests
the default 5600G configuration might be uncoupled or something.

Conclusion: Yes, 14800K is faster. But are you going to spend a lot of
            money for that much improvement ? You have to be careful
            about getting "trapped in incrementalism".

            A lot of lazy reviewers does not make this easy to evaluate.

            I have a scaling test for the 5950x, but the OS sucks at it.
            The study mostly shows strange scheduling choices. And Jim Keller
            already complained that another benchmark did not show off the
            best performance of his arch.

            That leaves us with Passmark 61379 versus 45780.
            A single point comparison, for your purchasing choice.

   Paul